In the near future, the Internet of things (IoT) will rapidly change and automate tasks in our everyday life. IoT networks have sensors measuring the environment and automated agents changing it with respect to predefined objectives. Modeling agents as web services requires lots of metadata from the environment in order to define the desired performance in a specific context. For this purpose, we propose an automatic measurement-based metadata creation method that analyses multivariate time series gathered from the sensors during agents change the environment. The time series analysis uses a cumulative sum algorithm (CuSum) to detect events and association rule learning to find temporal patterns. We evaluate our system with a Long-Term Evolution (LTE) simulator having mobile phones corresponding to IoT devices, LTE macro cells as the data source, and the Self-Organised Network (SON) functions as the automated agents in the network. Our experiments give promising results and show that the metadata creation process can be utilised to characterise IoT agents.